Interactive agents in multimodal settings repeatedly choose among a limited set of actions, labels, or ordinal outcomes. According to a new arXiv preprint, generating a free-form response for every such choice introduces unnecessary latency and makes the interface between model and environment more complex than it needs to be.
The paper, titled "Intern-Decision: Efficient Multimodal Decision Making with Autoregressive Masked Language Modeling," proposes an alternative that uses autoregressive masked language modeling to produce decisions more directly. The abstract argues that this can streamline the decision process, though the full method and experimental results are not detailed in the available text.
Because the source is limited to the abstract, the article does not include specific performance numbers or comparisons. The contribution appears to be a conceptual and architectural shift toward decision-specific outputs rather than general text generation, which could be relevant for robotics, dialogue systems, and other interactive AI applications.